GitLab

* CPU perf optimization in linear autograd function

Avoid enable_grad context when possible in cast function. Cache distributed group properties.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* CPU perf optimization in prepare_forward function

Avoid torch.nn.Module impl of __setattr__.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Avoid module import in TE module forwards
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use fast getter for params
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Reuse tensor dims in linear autograd func
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Apply optimizations to grouped linear
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug test failures
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Debug test failures
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Avoid deepcopy in tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move _fast_setattr logic to __setattr__ method
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

skip FP8 CP tests if hardware does not support FP8
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

Port optimizer tests to pytest
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* reduce atol/rtol for F16 tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* relax the tols for Ampere
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Add base class for tensor proxies
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move tensor detaching logic to tensor proxy base class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use Python wrappers to PyTorch extensions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Include transpose caching logic in proxy encode function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug dimension mismatch with amax history
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move dequantize logic to proxy_decode func
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename to "QuantizedTensor"
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename "proxy_detach" to "detach"
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Include transpose cache in detach and clone funcs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Update FP8 workspaces with QuantizedTensor functions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move logic for FP8 transpose cache in FP8 workspaces to base class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove cast-transpose logic from linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unnecessary args for Float8Tensor when using FP8 attr dict
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove __torch_function__ to QuantizedTensor
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update tests/pytorch/test_float8tensor.py
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Debug FP8 transpose test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug cast functions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* clean code for CP function args
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a placeholder for Ulysses implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* commit code change to CP+A2A
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* finish the draft fwd implementation of Ulysses
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add draft bwd implementation of Ulysses
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make swa work with ulysses
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* commit FP8 code for Ulysses
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix qkv type in the bwd of FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix qkv_dtype of FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code refactoring
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor code change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* config cp correction dtype of FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* code style change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* save chunk_ids
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* try to make Ulysses A2A async
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make more a2a async
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix a2a_outputs
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix chunk_ids generation for A2A
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* avoid code duplication of a2a before attn
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove code duplication of a2a after attn
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add cp_stream in A2A implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix qkv of fp8_fwd + bf16_bwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix kernel order in cp a2a communication
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning for CP a2a
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix merging with main
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix a2a communication order
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* adjust sequence chunk reordering for a2a
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add docstring for A2A implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change an assert info
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add unit tests of A2A implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add more A2A unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix CP unit tests
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add more cp unit tests
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix window size of no_mask
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fused attn does not support swa+no_mask
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change num_gqa_groups to 2 for A2A implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* function and variable renaming
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning for CP all-gather implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* some function renaming
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* commit code change for kv all-gather implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix all-gather implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a window size check
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add unit test of all_gather+no_mask
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix all-gather cp implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* code format fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code format fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix FP8 with A2A implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add paper references to CP implementations with all-gather and all-to-all
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change pdf to abs
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* elaborate cp_comm_type
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix CP docstring
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

---------
Signed-off-by: Xiaowei Ren <xren@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* propagate scale_inv modification to GroupedLinear
Signed-off-by: Xin Yao <xiny@nvidia.com>

* optimization for separate scale_inv of weights and single output
Signed-off-by: Xin Yao <xiny@nvidia.com>

* let grouped gemm support different input combinations
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix type
Signed-off-by: Xin Yao <xiny@nvidia.com>

* add contiguous check
Signed-off-by: Xin Yao <xiny@nvidia.com>

* use len() instead of isinstance
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix ut
Signed-off-by: Xin Yao <xiny@nvidia.com>

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* [TE/PyTorch][MoE] Add FP8 padding and unpadding module 

 1. Add multi-tensor padding kernel for FP8 with padding size = 16.
 2. Add FP8Padding and Fp8Unpadding module
 3. Add Padded GroupedLinear unit tests

---------
Signed-off-by: beinggod <zhangruibin@01.ai>
Co-authored-by: Phuong Nguyen <36155692+phu0ngng@users.noreply.github.com>

* fp8 mha with rope
Signed-off-by: Xin Yao <xiny@nvidia.com>

* avoid index select in cast ops
Signed-off-by: Xin Yao <xiny@nvidia.com>

* avoid index select in fused_attn_fwd
Signed-off-by: Xin Yao <xiny@nvidia.com>

* rename is_first_module_in_mha to fp8_output
Signed-off-by: Xin Yao <xiny@nvidia.com>

* resolve comments
Signed-off-by: Xin Yao <xiny@nvidia.com>

* resolve comments
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* move transpose to backward for fp8 input
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix ut
Signed-off-by: Xin Yao <xiny@nvidia.com>

* resolve comments
Signed-off-by: Xin Yao <xiny@nvidia.com>

* update argument list for CP
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix for FA3
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* remove unnecessary copy of scale_inv
Signed-off-by: Xin Yao <xiny@nvidia.com>

* skip fp8 dpa/mha tests when fa3 is not available
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix a merge bug
Signed-off-by: Xin Yao <xiny@nvidia.com>

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* fix FP8 logic when FA3 is not installed
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor tweak to make logic more explicit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* limit FA3 warning to Hopper and NVTE_FLASH_ATTN=1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* prefer fused attn for FP8
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove dtype from args
* update docs with permutation ops

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>

* WIP: add fa3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: clean up
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* WIP: add benchmarks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* differentiate func/varlen_func
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix parsing keyword for FA3 and remove bshd->thd conversion for flash_attn_func
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: add FP8 fwd support
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* add FA3 FP8 fwd code and test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix assert for FA3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix FA3 FP8 logic and add tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* update FA2 to <=2.6.3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* tweak unit tests for base/mask
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* set constraints for FA3 for sm90 and causal_bottom_right
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert debug changes in benchmark script
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add permutation functions

* Add permutation ops

* Remove the dependency on cutlass

* Move permutation.py out of module dir

* Rewrite the unit test and enable skipping if FP8 is unavailable

* Rename exposed C++ API and reorder its parameters + take NVTETensor as inputs

* Use Float8Tensor for FP8 input

* Move dtype to ctx

---------
Signed-off-by: Jiang Shao <jiangs@nvidia.com>
Co-authored-by: Qi Zhang <qizhang@nvidia.com>
Co-authored-by: Phuong Nguyen <36155692+phu0ngng@users.noreply.github.com>

* add support for padding in UnfusedDPA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* add support for padding_causal/_bottom_right
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix padding_causal/_bottom_right
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* need to test max512 backend
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* revert last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix mask logic in unfused
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* use actual_seqlen for alibi/causal_bottom_right padding
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes and convert causal to causal_bottom_right for inference
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* use causal in kv cache inference test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* simplify get_alibi logic
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* simplify the non-padding path for get_alibi
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* avoid batch_size loop in generating padding_causal/_bottom_right masks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* add window_size to AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add seq_offsets_qkvo for cudnn thd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add seq_offsets_qkvo to AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix seq_offsets calculation of cudnn thd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove a thd assert
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix bias for thd test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add thd test for cudnn FA with CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* skip GQA/MQA test for cuDNN THD
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make sure seq_offsets are computed with qkv_group of hd_hd_hd while CP>1
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix seq_offsets inputs
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove two comments
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attn mask type for cudnn thd with cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attn_mask_type check
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attn_mask_type for cudnn fa with thd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix a typo
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix out dout in bwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert cudnn+thd does not support attn bias
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* check if attn_mask_type has padding
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change cp test batch size to 2
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix code format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix two assert info
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert comment
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert comments
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert comments
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert swa+CP cannot work with thd format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a new CP function for swa
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a missing dgrads
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add draft fwd function for swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* enable flash attention for swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove an assert of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* call SWAFuncWithCP for swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* use 2hd layout
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* change qkv_format check
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a code comment
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* tensor shape bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* tensor shape fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add function to compute cu_seqlens of a cp rank
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add cu_seqlens and cu_seqlens_padded to context parallelism
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix FlashAttention output sequence length
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix cu_seqlens_kv_per_step calculation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero dQKV for ending padded tokens
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero dQKV tensors of FlashAttention
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix softmax_lse correction
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove padded tokens of KV to save comounication
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* do not need to zero dkv for FlashAttention any mroe
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero out tensors
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix CP unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix kv shape of cp test with thd format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* update cp unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* add simple code framework
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* try not to have a separate CP function for SWA
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* backup some code change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* back up code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* clean up fwd implementation of SWAFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert info
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* reduce kv chunk concat overheads
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make AttnFuncWithCP and SWAFuncWithCP have same API
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a docstring
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* preliminary implementation of SWAFuncWithCP forward seems working
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix output shape of SWAFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code refactoring for FlashAttention and add a code placeholder for bwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* use gather_along_first_dim
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* finish the preliminary implementation of bwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert condition
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add draft implementation of SWA+CP with FusedAttention
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attention mask type of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add qkv_layout
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add missing window_size argument
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix kv shape of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug and typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix dout shape
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add multi stream in fwd of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* save chunk_ids_to_kv_ag in fwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add multi stream in bwd of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor fix to cp stream sync
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* rename AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* check if window size is None
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix docstring of AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add env var for users to choose KV ag or KV p2p
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* update cp tests
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix window size in cp unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix pytest skip messages
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add cp_comm_type into API
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* add deterministic konb in cuDNN fused attn backend
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* pass fp8 and fp8_meta to attn_func_with_cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert only Fused Attn can support FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant assert
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a fwd draft implementation of FP8 + CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* save fp8 and fp8_meta
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert sequence length divisible requirements
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* remove a redundant qkv_layout compute
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* if condition change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* some typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add support table of context parallelism
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* typo and code format fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* do not print multiple disabling messages
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix aux_ctx_tensors of FP8
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix device in torch.arange and adjust code for the PR of MLA
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* commit code change for FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* commit more code change for FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* commit more fp8 code for FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fixes
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* cast merged CP results from FP32 to BF16
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix softmax_lse
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix some bugs of FP8 dkv exchange
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add FP8 unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix typos and clean asserts
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix get_p2p_comm_info
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix dkv p2p exchange
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change FP8 dkv P2P to A2A
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add FP8+CP unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert amax reduction is needed for FP8+CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* remove duplicated code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* destroy process group in CP unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove interval from fp8_recipe because it has been deprecated
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* try to fix the failed CP test with the latest CI pipeline
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* remove redundant f before string
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change META_O_CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

---------
Signed-off-by: Xiaowei Ren <xren@nvidia.com>
Co-authored-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xiaowei Ren <xren@cs-cw-dfw-login-01.cm.cluster>

feat(pytorch): Allow TransformerLayer and MultiheadAttention to accept sequence length parameters (#1066)

* Added ability for seqlen for transformer and mha layer
Signed-off-by: Lukasz Pierscieniewski <lukaszp@nvidia.com>

* Documentation for new parameters
Signed-off-by: Lukasz Pierscieniewski <lukaszp@nvidia.com>

* Add tests for THD layout, assert for THD layout with KV-Cache
Signed-off-by: Lukasz Pierscieniewski <lukaszp@nvidia.com>

* Fixed tests
Signed-off-by: Lukasz Pierscieniewski <lukaszp@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Move THD logic in shape calculation, add missing optional in params
Signed-off-by: Lukasz Pierscieniewski <lukaszp@nvidia.com>

* Skip the THD test on GPUs older than Ampere
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

---------
Signed-off-by: Lukasz Pierscieniewski <lukaszp@nvidia.com>
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Przemek Tredak <ptredak@nvidia.com>

* support dtype casting fusion in FusedAdam
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* minor changes
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* fix lint
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* changes based on review comments
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* remove unused code
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* code refactor
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* fix typo
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* refactor
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* remove unused code
Signed-off-by: Shijie Wang <jaywan@nvidia.com>

* Fix linter warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Copy CUDA headers for framework sdists
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Shijie Wang <jaywan@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

* add window_size to AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add seq_offsets_qkvo for cudnn thd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add seq_offsets_qkvo to AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix seq_offsets calculation of cudnn thd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove a thd assert
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix bias for thd test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add thd test for cudnn FA with CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* skip GQA/MQA test for cuDNN THD
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make sure seq_offsets are computed with qkv_group of hd_hd_hd while CP>1
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix seq_offsets inputs
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove two comments
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attn mask type for cudnn thd with cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attn_mask_type check
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attn_mask_type for cudnn fa with thd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix a typo
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix out dout in bwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert cudnn+thd does not support attn bias
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* check if attn_mask_type has padding
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change cp test batch size to 2
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix code format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix two assert info
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert comment
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert comments
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert comments
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert swa+CP cannot work with thd format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a new CP function for swa
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a missing dgrads
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add draft fwd function for swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* enable flash attention for swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove an assert of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* call SWAFuncWithCP for swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* use 2hd layout
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* change qkv_format check
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a code comment
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* tensor shape bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* tensor shape fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add function to compute cu_seqlens of a cp rank
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add cu_seqlens and cu_seqlens_padded to context parallelism
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix FlashAttention output sequence length
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix cu_seqlens_kv_per_step calculation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero dQKV for ending padded tokens
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero dQKV tensors of FlashAttention
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix softmax_lse correction
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove padded tokens of KV to save comounication
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* do not need to zero dkv for FlashAttention any mroe
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero out tensors
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix CP unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix kv shape of cp test with thd format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* update cp unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* add simple code framework
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* try not to have a separate CP function for SWA
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* backup some code change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* back up code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* clean up fwd implementation of SWAFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert info
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* reduce kv chunk concat overheads
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make AttnFuncWithCP and SWAFuncWithCP have same API
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a docstring
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* preliminary implementation of SWAFuncWithCP forward seems working
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix output shape of SWAFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code refactoring for FlashAttention and add a code placeholder for bwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* use gather_along_first_dim
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* finish the preliminary implementation of bwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix assert condition
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add draft implementation of SWA+CP with FusedAttention
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix attention mask type of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add qkv_layout
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add missing window_size argument
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix kv shape of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug and typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix dout shape
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add multi stream in fwd of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* save chunk_ids_to_kv_ag in fwd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add multi stream in bwd of swa+cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor fix to cp stream sync
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* rename AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* check if window size is None
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix docstring of AttnFuncWithCP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add env var for users to choose KV ag or KV p2p
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* update cp tests
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix window size in cp unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix pytest skip messages
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add cp_comm_type into API
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* code cleaning
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* assert sequence length divisible requirements
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* add support table of context parallelism
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* typo and code format fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* do not print multiple disabling messages
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix device in torch.arange and adjust code for the PR of MLA
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix typos and clean asserts
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

---------
Signed-off-by: Xiaowei Ren <xren@nvidia.com>
Co-authored-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xiaowei Ren <xren@cs-cw-dfw-login-01.cm.cluster>

* update example/benchmark scripts
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix head_dim after MLA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update notebook
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* merge k_channels and v_channels back to kv_channels and accept a tuple
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix isinstance call
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix MLA tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add op for in-place add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add op for in-place add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add op that adds extra output to fuser
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add fused op for GEMM+bias+add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add fused op for dgrad+add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add documentation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Fix linter warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Review suggestions from @ptrendx

Output tensor dtype and device take precedence over weight tensor in linear functional API. Move some index calculation to fuser constructor. Avoid some unnecessary dereferences.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug test failures
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update transformer_engine/pytorch/ops/fuser.py
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* use fused_multi_cast_transpose
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix input being empty tensor
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* allocate output tensors in C++
Signed-off-by: Xin Yao <xiny@nvidia.com>

* simplify code
Signed-off-by: Xin Yao <xiny@nvidia.com>

* avoid cudaGetDriverEntryPoint
Signed-off-by: Xin Yao <xiny@nvidia.com>

* reduce torch.Tensor() calls
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* update test
Signed-off-by: Xin Yao <xiny@nvidia.com>

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

[C/PyTorch] Fixed incorrect use of `torch.distributed.new_group()` when creating intra-node group in `initialize_ub()` (#1087)

* updated initialize_ub() to use new_subgroups_by_enumeration() to generate intra-node groups, added new unit tests for TE layers with comm overlap
Signed-off-by: Alp Dener <adener@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* add multi-latent attention for DPA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix Jax/Paddle API
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix typo in test script
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix too-many-boolean lint error
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "fix lint"

This reverts commit 67399a3a6f45bb4ce9e5eaa6bcce40b28e347e5b.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix stride check in get_qkv_layout
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: fix layout_thd tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* WIP: debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix merge conflict
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix thd pad_between_seqs=False/True tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* use 2hd layout
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* change qkv_format check
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add a code comment
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* tensor shape bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* tensor shape fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add function to compute cu_seqlens of a cp rank
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add cu_seqlens and cu_seqlens_padded to context parallelism
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix FlashAttention output sequence length
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix cu_seqlens_kv_per_step calculation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero dQKV for ending padded tokens
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero dQKV tensors of FlashAttention
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix softmax_lse correction
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove padded tokens of KV to save comounication
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* do not need to zero dkv for FlashAttention any mroe
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* zero out tensors
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix CP unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix kv shape of cp test with thd format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* update cp unit test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* remove redundant code
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

---------
Signed-off-by: Xiaowei Ren <xren@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: Xiaowei Ren <xren@cs-cw-dfw-login-01.cm.cluster>

* fix workspaces and unfused bias in multi-stream cuBLAS

* Expose num_streams via pybind

* Fix C-compatibility

* rm importing packaging in test_fused_attn.py

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Ensure that the inputs to custom calls are contiguous
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Fixes
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Added test
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Fixes
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Fixes from review
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix typo
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Rm unused import causing CI failures
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

[PyTorch] Fixing hang in `initialize_ub()` for multi-node runs after PR901 removal of MPI-dependence (#986)

* Re-implementing PR901 (removing MPI-dependence in Userbuffers) with multi-node fixes

* passing data-parallel rank/size info from torch.distributed to userbuffers
Signed-off-by: Alp Dener <adener@nvidia.com>

* multi-node example working with UB_SKIPMC=1 but not with multicast
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed multi-node hang in initialize_ub(), updated comm+GEMM overlap example to support multi-node mixed tensor/data parallelism, added README
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed use case when Userbuffers is asked to allocate the TP overlap buffer with UB_SKIPMC=1
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected example problem to set device by local ordinal instead of global process rank
Signed-off-by: Alp Dener <adener@nvidia.com>

* double-free fix in userbuffers destructor
Signed-off-by: Alp Dener <adener@nvidia.com>

* removed unnecessary and incorrect torch.cuda.set_device(...)
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected inter-node ranks logic
Signed-off-by: Alp Dener <adener@nvidia.com>

* generalized node ID logic in initialize_ub to handle arbitrary world rank layouts within node
Signed-off-by: Alp Dener <adener@nvidia.com>

* added single-node comm+GEMM overlap unit tests
Signed-off-by: Alp Dener <adener@nvidia.com>

* LayerNormMLP example confirmed working with 2 nodes on Eos
Signed-off-by: Alp Dener <adener@nvidia.com>

* unit test cleanup
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected DP group ranks logic in LNMLP comm+GEMM overlap example
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected enums in unit test
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed incorrect Ubuf object init signature
Signed-off-by: Alp Dener <adener@nvidia.com>

* switched default backend for Userbuffer bootstrapping to Gloo with MPI and NCCL fallbacks, and initialize_ub option to manually select backend
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed all comm+GEMM overlap unit tests
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected all_gather use for Gloo backend
Signed-off-by: Alp Dener <adener@nvidia.com>

* changed userbuffers allgather callback to always use all_gather() instead of all_gather_into_tensor()
Signed-off-by: Alp Dener <adener@nvidia.com>

* restored and verified old MPI-based bootstrapping via NVTE_UB_WITH_MPI=1 option at compile time
Signed-off-by: Alp Dener <adener@nvidia.com>

* disabled scoped GIL release for comm+GEMM overlap algorithms
Signed-off-by: Alp Dener <adener@nvidia.com>

* avoid dist.init_device_mesh in comm+GEMM overlap example to support older PyTorch versions
Signed-off-by: Alp Dener <adener@nvidia.com>

* applied RS overlap FP8 fix from PR1004
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed segfault in Userbuffers destructor
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected comm+GEMM overlap unit test arguments
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed unit test run command for when Userbuffers is compiled with MPI
Signed-off-by: Alp Dener <adener@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Refactored torch.distributed collectives into pure C++ callbacks
Signed-off-by: Alp Dener <adener@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add option to pass kwargs to CUDA graph module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Debug unit tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Tweak comments
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* add cuDNN swa
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix SWA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add set_deterministic and minor fixes for swa
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add AttentionParams
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* change window_size to int64_t; fix swa/determinism tests; cache _attention_backends
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add window_size to get_backend; fix jax and paddle
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes; add set_deter to bwd_impl
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix unit tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix FP8 tests due to determinism
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add support matrix for SWA and bias
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes and lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add wording on window_size special cases
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor tweak on wording
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix jax assertion error
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix wording
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* call bwd with deterministic=true for jax/paddle
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add determinism words in documentation
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add basic infrastructure for Sequential module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 support in linear op

Runs, but need to validate. Runtime errors with non-FP8 params and FP8 compute, or FP8 params and non-FP8 compute.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add reshape op and unit test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add unfused linear op

Test does not pass with FP8.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug unfused linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add test for linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add separate abstract classes for unfused and fused ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Consolidate unfused ops in submodule
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add linear-bias fused op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use fused cast-transpose in linear ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable GEMM+bias fusion with FP32 activations

Not supported by cuBLAS.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add parallel unit test for unfused linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Refactor parallel tests to reduce job launches
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add all-reduce, all-gather, and reduce-scatter ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unused file
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug multi-GPU FP8 test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add support for FP8 scale updates

Still need to implement amax reductions.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add license boilerplate
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fuse GEMM+bias in row TP

Add documentation for unfused ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename pipeline to fuser

Expand documentation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Tweak documentation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Preserve cached FP8 transpose between ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add option for fused wgrad accumulation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Directly output FP8 from linear if needed
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix cuDNN front-end commit
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use updated FP8 tensor API for transpose caching
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use updated API for FP8 scale updates
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add tests for non-default FP8 recipes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename UnfusedOperation to BasicOperation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add unit test to check amax reduction with fusable op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Operator autograd state no longer needs to be initialized
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Initial functional implementation of linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug fused linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove autograd context from functional linear impl
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use functional linear impl in fused linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename subdirectory from "fuser" to "ops"

Avoid confusion with kernel fusers and graph compilers.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update with Float8Tensor changes in #820
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unnecessary CPU overheads
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Correctly pass FP8 metadata from next op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter errors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add convenience functions to manipulate Sequential class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Clear saved tensor data in linear op after bprop
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix Pylint error
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix test name in QA script
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Run distributed tests even when only 1 GPU is available
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Only run distributed tests with 2 GPUs if there are >=2 GPUs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Review suggestions from @sudhakarsingh27 and @ksivaman

Fix spelling of "fusible". Avoid "input" name in internal APIs.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update transformer_engine/pytorch/ops/__init__.py
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* update to FE 1.5.1 and add bottom right causal
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* adjust logic for backend selection
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* update FE to 1.5.2
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add get_attention_backend function
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* update get_attention_backend
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix get_attention_backend
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* tweak get_attention_backend and fix unit tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes for unfused, get_backend, etc
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Update transformer_engine/pytorch/attention.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix cpu offload
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes for get_attention_backend
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* explicitly skip FP32 and padding tests because there is no support
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fix for window size check
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update check_set_window_size and add enc_dec_attn_mask_type/enc_dec_window_size
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* update FE to 1.5.2
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* enable unfused attn for cross attn
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* unify logging info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* omit cudnn 9.1.1 and 9.2.1 due to bugs
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* set cu_seqlens_padded to cu_seqlens by default
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace variable name with ctx.variable
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "enable unfused attn for cross attn"

This reverts commit bc49f14fca904217a711b4a86c45a4a739a17a14.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* restrict cudnn version for fp8 tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove mha_fill for FP8
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "remove mha_fill for FP8"

This reverts commit 83ffc44114dc6eb3d426d742b6c5a4d34805ec04.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* lower cudnn version to >=9.2.1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* disable CP-THD tests for fused attn on <sm90
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

cache was added in Python 3.9.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* GroupedGEMM via multi-stream cublas

* fix A/B is nullptr while D is not nullptr

* add fp8 grouped gemm

* register with TorchScript

* add the GroupedLinear layer

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Jiang Shao <jiangs@nvidia.com>
Co-authored-by: Qi Zhang <qizhang@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* simplify offset tensors
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes; tests pass
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix C lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace with_offset with with_padding
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace with_padding with padded
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes after merge
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fix for fused attn fwd/bwd calls
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix Jax
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* adjust spacing in docstring
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix pytorch tests; fix paddle api
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix attn_biases
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix AttnFuncWithCP backward
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix jax
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix attn with CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix paddle
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* subclass DPA with BaseModule and test with test_gpt_checkpointing
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* test DPA only
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* test save and load
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor tweaks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor tweak
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add hook in case core_attention._extra_state is missing
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* check named buffers in BaseModule; remove FP8 scratchpad override function; test FP8 for sm90+
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes: test size, interval in recipe, named_buffer loop
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* move BaseModule from FusedAttention to DPA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Apply formatting
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Apply formatting
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>